Variable declared through structured-bindings feature not visible to lambda capture

[llvmbot]

Bugzilla Link	29161
Resolution	INVALID
Resolved on	Jan 30, 2018 16:33
Version	4.0
OS	Linux
Blocks	#35152
Reporter	LLVM Bugzilla Contributor
CC	@bebuch,@zmodem,@zygoloid,@HighCommander4

Extended Description

Code sample:

struct P { int i = 1; int j = 2; };

int p(int i) { return i; }

int main()
{
auto [ i , j ] = P();

p(i); // works

i { return i; }(); // does not work

return 0;
}

Defining a lambda in the "does not work" line results in the error message below:

test.cpp:11:4: error: 'i' in capture list does not name a variable
i { return i; }(); // does not work

Somehow i is not accessible from the lambda capture if it is declared through structured bindings. There is no problem when declaring it differently, e.g. as "auto i = P().i;"

Versions used:
LLVM 4.0 (head), last commmit a733d39
Clang 4.0 (head), last commit 4fef27d8c527b0da6135831dbb6cbe45565b5d44

Compiler options: std=c++1z -stdlib=libc++ -O0

[llvmbot]

Interestingly enough

auto [ i , j ] = P();

x = i { return x; }();

This variant compiles fine because the bug seems to be in the branch for a lambda capture withOUT an initializer:

https://reviews.llvm.org/diffusion/L/browse/cfe/trunk/lib/Sema/SemaLambda.cpp;306567$1058

  Var = R.getAsSingle<VarDecl>();

template
DeclClass *getAsSingle() const {
if (getResultKind() != Found) return nullptr;
return dyn_cast(getFoundDecl());
}

In the case of structured binding, getFoundDecl() seems to return BindingDecl <: ValueDecl whereas the call to getAsSingle() expects VarDecl <: DeclaratorDecl <: ValueDecl.

[zygoloid]

In the example in comment#0, i and j are not variables, and as such are not subject to lambda capture. The issue is with the c++ language rules, not with Clang. Sorry.

[bebuch]

I'm quiet sure this is not correct, i and j are variables created by structured binding initialization. (C++17)

http://en.cppreference.com/w/cpp/language/structured_binding

Here is a shorter version of the test case:

int main(){
int a[1] = {0}; // array, struct or tuple-like type with one element
auto [b] = a; // structured binding initialization
[b]{}; // lambda capture
}

$ clang++ -std=c++17 main.cpp
main.cpp:4:6: error: 'b' in capture list does not name a variable
[b]{};
^
main.cpp:4:5: warning: expression result unused [-Wunused-value]
[b]{};
^~~~~
1 warning and 1 error generated.

$ clang++ --version
clang version 7.0.0 (http://llvm.org/git/clang.git aff1a3f56c049bb08f51d074eeecd591118f7cc3) (http://llvm.org/git/llvm.git 20b5be1)
Target: x86_64-unknown-linux-gnu
Thread model: posix

Code is accepted by GCC in C++17 mode.

[zygoloid]

Sorry, GCC is wrong here. b is a structured binding, not a variable. And cppreference is not an authoritative source.

Permitting lambda capture of structured bindings was intentionally deferred by the c++ committee because it was unclear what that would mean (do you capture the complete underlying object or just the part named by the binding?) and there wasn't time to resolve that question prior to the publication of C++17.

[bebuch]

Okay, thanks a lot for the explanation! I didn't know that a structured binding isn't a variable itself.

Then Bug 34749 is also invalid, right?

[zygoloid]

*** Bug llvm/llvm-bugzilla-archive#34749 has been marked as a duplicate of this bug. ***

[zygoloid]

For clarity: I would guess that the C++ committee will choose to allow lambda capture of structured bindings at some point in the future. I wouldn't want to guess whether it will be considered a defect resolution against C++17 or a C++20 feature, nor what the precise semantics will be (does a capture by copy capture the complete object or just the bound portion?), but Clang will of course implement that behavior in the relevant modes once it is decided upon by the committee.

[zygoloid]

mentioned in issue llvm/llvm-bugzilla-archive#34749

[zmodem]

mentioned in issue #35152